Imas Sukaesih Sitanggang

Peatland fires in Southeast Asia are strongly influenced by hydrological conditions, yet most studies rely on low-temporal-resolution data, limiting the capture of rapid hydrological dynamics. This study analyzes high-frequency hydrological dynamics in relation to fire occurrence in protected peatlands of Central Sumatra. Real-time monitoring using 12 IoT sensors recorded groundwater level (GWL), soil moisture (SM), and soil temperature (ST). Results show substantial spatial variation, with GWL (−144.30 to 9.00 cm), SM (14.52% to 94.82%), and ST (23.01 to 33.93 °C). Linear mixed-effects models (LMMs) and first-order autoregressive (AR (1)) temporal correlation showed significant associations of GWL with SM (β = 0.460, p < 0.001) and ST (β = −0.516, p < 0.001). GWL decreased by 13.73 cm (−0.54 cm day⁻¹) during the pre-fire period and by a further 6.34 cm (−0.47 cm day⁻¹) during the fire period (27 August 2024–9 September 2024), when the descriptive mean GWL reached −89.43 cm. Post-fire, GWL increased by 41.99 cm (0.82 cm day⁻¹), while SM increased and ST decreased, coinciding with reduced hotspot activity. The fire period was identified using SIPONGI hotspots and SIPP fire records and suppression. These findings show temporal associations between hydrological dynamics and fire occurrence, supporting high-frequency IoT monitoring of peatland hydrology.

Background: The rapid growth of online culinary platforms has produced a vast amount of unstructured Indonesian recipe texts. Extracting structured information such as ingredients, quantities, and measurement units002C is essential for downstream applications, including nutritional computation and diabetes-oriented dietary recommendation systems. However, Indonesian recipe texts present substantial challenges due to linguistic variability, informal writing styles, and limited availability of annotated corpora. Objective: This study investigates effective modeling strategies for Indonesian recipe Named Entity Recognition (NER) in low-resource settings and to identify reliable approaches to improve the robustness and generalization of extraction. Methods: A quantitative experimental approach was employed to evaluate the IndoBERT, BiLSTM-CRF, IndoBERT-CRF, and hybrid IndoBERT-BiLSTM-CRF architectures. A reliability-weighted voting ensemble combining heterogeneous models was proposed, and high-confidence ensemble predictions were used for pseudo-labeling in a semi-supervised setting. Performance was assessed using entity-level precision, recall, and F1-score. Results: IndoBERT achieved a strong baseline performance (F1-score = 96.61%). IndoBERT-CRF and the hybrid IndoBERT-BiLSTM-CRF model achieved competitive F1-scores (92.99% and 92.82%, respectively) but did not outperform the Transformer-only baseline. The proposed weighted ensemble achieved the best supervised performance (F1-score = 96.79%). Semi-supervised training with ensemble-generated pseudo-labels maintained competitive performance (F1-score = 96.72%) while improving recall and generalization to unseen ingredient expressions. Conclusion: Increased architectural complexity does not necessarily improve performance on the Indonesian recipe NER task. Reliability-weighted ensemble learning combined with pseudo-labeling provides a stable and effective approach for low-resource culinary NER and supports future nutrition-oriented recipe processing. Keywords: Indonesian Recipe Text; Pseudo-Labeling; Named Entity Recognition; Semi-Supervised Learning; Weighted Ensemble

Visual Question Answering (VQA), a multimodal AI task combining computer vision and language understanding, answers text-based questions about visual content. Although VQA has developed rapidly, most existing approaches assume a single correct answer, whereas real-world images often require several valid answers simultaneously. This study proposes MangoVQA, to our knowledge the first multi-answer VQA framework for mango quality grading aligned with SNI 3164:2024. MangoVQA separates perception from reasoning, combining YOLOv11n-based neural perception with deterministic, rule-based question interpretation, unlike end-to-end learned-fusion VQA models. Each answer is grounded in detected mango instances, enabling traceable multi-answer responses rather than opaque predictions. The YOLOv11n detector localizes fruits and identifies four quality grades (Extra Class, Class I, Class II, and Reject), while the Perspective Engine extracts color and defect attributes through calibrated HSV analysis. These facts populate a structured answer representation, queried via predefined intents and slots. From 468 annotated images, the pipeline yielded 17,764 question-answer pairs across 22 question types, two reasoning levels, and two languages. 22.85% of questions required multi-answer responses, scored with set-based precision, recall, F1-score, and subset exact match. Evaluation reaches 0.919 accuracy on global open-ended questions, 0.937 on local open-ended questions, and 0.846 micro-F1 on global multi-answer questions, exceeding the 62.91% F1 reported by EMQA, the only prior multi-answer VQA baseline, on a different dataset. Practically, MangoVQA offers auditable decision support for smallholders, cooperatives, exporters needing SNI-compliant evidence, and mango customers; theoretically, decoupling visual perception from text interpretation and reasoning offers a viable, interpretable alternative to end-to-end multimodal fusion.

Visual question answering (VQA) is a challenging research area that enables machines to answer natural language questions based on visual content by jointly understanding images and text. Conventional VQA systems typically produce a single answer for each image–question pair. However, many real world visual questions are ambiguous or complex, allowing multiple valid answers to exist. This systematic literature review (SLR) focuses on multi answer VQA systems and the use of object detection, following the PRISMA 2020 guidelines. We analyzed 58 peer-reviewed journal articles retrieved from the Scopus database published between 2020 and 2025. Ten of these studies clearly stated that generating multiple answers was their main goal. Forty-eight others indirectly supported answer variability by using object-based or multi-instance reasoning. Through this review, we examine the current methodologies for supporting multi-answer generation, including model architecture, datasets, and evaluation metrics. Most multi answer generation approaches utilize attention mechanisms, graph neural networks, and transformer-based models. Additionally, we propose a taxonomy of multi-answer VQA organized along four dimensions. Limitations are identified in datasets and evaluation metrics (i.e., answer ambiguity/subjectivity). Future research should focus on improving model interpretability and designing an evaluation framework that incorporates subjective and context-sensitive responses.

PM2.5 concentrations in Jakarta frequently exceed World Health Organization (WHO) air quality guidelines, indicating the need for an integrated approach for pollution prediction and source assessment. This study develops a spatiotemporal prediction framework using a long short term memory (LSTM) model integrated with the hybrid single particle Lagrangian integrated trajectory (HYSPLIT) model for backward trajectory analysis. Daily PM2.5 data from five monitoring stations were combined with meteorological variables from ERA5, Visualcrossing, and the global data assimilation system, with spatial context evaluated using Sentinel-2 land cover maps. After hyperparameter tuning, the optimized model demonstrated robust predictive capabilities, achieving a peak coefficient of determination (R2) of 75.87% on the test data. The framework exhibited exceptional relative accuracy, particularly at the Jagakarsa and Kebun Jeruk stations, which recorded mean absolute percentage error (MAPE) values of 13.34% and 17.80%, respectively. Backward trajectory analysis during selected pollution episodes indicates two dominant regional transport pathways that may influence PM2.5 levels in Jakarta. These pathways are associated with air mass transport over industrial and built-up areas in eastern and northern regions surrounding Jakarta. Land cover analysis shows limited vegetation along these pathways. Overall, elevated PM2.5 events are associated with combined local emissions, regional transport, and meteorological conditions that limit pollutant dispersion near the surface.

Stocks are a popular investment instrument but carry high risks, where investors may incur losses when stocks are bought at high prices and sold at lower prices. Technical analysis is used to study past stock price behavior to predict future prices. In this study, grammatical evolution (GE) is applied as an evolutionary computing technique to discover optimal functions or programs that represent historical stock price data. This study develops GE based prediction models by utilizing objective functions and search spaces defined through grammar. The model integrates technical indicators based on complex statistical models such as autoregressive integrated moving average (ARIMA), prophet, exponential smoothing, and Fibonacci retracements. Furthermore, this study employs GE to generate ensemble weights randomly, ensuring each model contributes equitably to the final prediction formula. Experiments were conducted using multiple stock datasets, including SMAR, S&P 500, the Johannesburg Stock Exchange (JSE), the New York Stock Exchange (NYSE), and Adani Enterprises (ADANIENT), to evaluate the model’s adaptability and generalization capability. The results demonstrate that the proposed GE model effectively captures complex market patterns and produces more reliable stock price predictions compared to deep learning-based approaches. Although GE requires greater computational time, the findings suggest that GE provides a flexible and effective framework for constructing hybrid stock price forecasting models in dynamic market environments.

Forest and land fires pose significant environmental challenges, causing economic and ecological damage depending on their severity. This study proposes a deep learning-based classification model to assess fire severity using the MobileNetV2 architecture. A dataset of 560 post-fire images was categorized into five severity levels, with dataset preprocessing involving resizing, rescaling, and image augmentation. To enhance model performance, K-means clustering was applied for balanced data distribution across classes. The model was trained using grid search for hyperparameter tuning, with the optimal combination being a batch size of 8, learning rate of 0.0001, and dropout of 0.3. Training was conducted in 50 epochs, and evaluation using the confusion matrix demonstrated an accuracy of 85%, precision of 86%, and recall of 81%. The results indicate that MobileNetV2 effectively classifies post-fire severity levels, offering a reliable tool for post-disaster assessment. This study highlights the significance of dataset preprocessing and hyperparameter tuning in improving model accuracy. Future research should explore alternative architectures and expand the dataset to enhance model generalization. These findings can aid authorities in assessing fire impact, supporting mitigation strategies, and improving post-fire land management.

The severity of forest and land fires is a crucial indicator for assessing their impact on ecosystems, particularly vegetation and soil. The assessment results serve as a foundation for forest and land restoration, rehabilitation, and conservation efforts. This study employs a deep learning algorithm to develop a forest and land fire severity assessment module. The CNN model used is MobileNetV2 that has an accuracy of 88.8%. The smart module is integrated into the Indonesian Forest and Land Fire Prevention Patrol Mobile Application and follows the Software Development Life Cycle approach in its development. Field observation images are input to the CNN module in the mobile application. The module then analyzes the fire severity and classifies it into very light, light, moderate, severe, and very severe categories. Testing results indicate that the module accurately predicts fire severity based on established assessment standards. The optimal time for capturing images is a few days after the fire, during daylight hours, to ensure the majority of images depict burned areas. Additionally, the findings highlight that lighting conditions and image quality significantly influence the accuracy of severity predictions. Further development is required to enhance the module's compatibility and flexibility, enabling its use across various devices.

This study presents a comparative analysis of 7 Convolutional Neural Network (CNN) architectures—MobileNetV2, VGG16, VGG19, LeNet5, AlexNet, ResNet50, and InceptionV3—for classifying post-forest fire areas using field-based vegetation imagery. A total of 56 models were evaluated through combinations of batch size, input size, and optimizer. The results show that MobileNetV2, VGG16, and VGG19 outperformed other models, with validation accuracies exceeding 88%. MobileNetV2 emerged as the most balanced model, achieving 96% accuracy with the lowest model size and training time, making it ideal for resource-constrained applications. This study highlights the potential of CNN-based classification using mobile field imagery, offering an efficient alternative to costly and condition-dependent satellite or drone data. The findings support real-time, localized identification of burned areas after forest fires, providing actionable insights for prioritizing recovery areas and guiding ecological restoration and land rehabilitation strategies.

Bulukumba Regency, located along the Walanae Fault and within a seismic gap zone, indicates the potential for future earthquake recurrence. However, the regional and community capacity to address earthquake hazards remains weak, as evidenced by the lack of regulations accommodating earthquake studies in Bulukumba. This study aimed to design an earthquake mitigation model based on a stakeholder approach in Bulukumba Regency. The methodology employed MACTOR (Matrix of Alternatives for Choice and Trade-Offs), utilizing survey and questionnaire data. The output is a framework for policymakers in earthquake mitigation activities. The results suggested two effective alternative models: (i) a stakeholder formulation model based on role capacity and (ii) a time segmentation model for stakeholder involvement in earthquake mitigation. Based on these two models, it is essential to establish strong coordination and collaboration among these actors in order to minimize the impact of disasters on both the community and the environment.